New server chips carry hidden cost


In the rush to roll out servers that put multiple processors on a chip, IBM, Intel, Sun and others inadvertently are raising software costs for users.

In the rush to roll out servers that put multiple processors on a chip, IBM, Intel, Sun and others inadvertently are raising software costs for users.

At issue is that software vendors such as Oracle and Microsoft that license software on a per-CPU basis are likely to consider each processor a separate CPU, a practice that means double the licensing costs for enterprise users deploying new dual-core servers. These servers have two processors on a single piece of silicon. With four-core servers and more on the horizon, the issue likely will become amplified as vendors try to figure out how to price software and enterprise users work to manage IT budgets.

At the same time, other technologies - such as server virtualization and grid computing - that fall under the utility computing umbrella, in which computer resources are pooled and workloads shifted according to application demands, also are causing some tricky software licensing issues.

The trend toward multi-core systems is an effort by server chip makers to drive up performance, while reining in system power requirements and heat generation. By using these architectures, vendors can use more lower-power cores - which generate less heat - instead of ramping up the frequency on single cores to boost performance. IBM has had a dual-core chip since 2001, and Sun and HP rolled out dual-core processors earlier this year. Intel and Advanced Micro Devices recently announced that they would move to dual-core designs in 2005.

"Multi-core is a symptom of a bigger issue," says Al Gillen, research director of system software at IDC. "The bigger issue is that we have changing technologies that are making historical licensing models no longer directly compatible with the way people are buying and deploying their equipment today."

W.L. Gore & Associates, best known for its Gore-tex fabric, negotiated a software license with Oracle late last year and then began updating its Sun hardware. In the process of bringing in servers based on Sun's first dual-core chip, UltraSparc IV, the Newark, Del., company discovered that Oracle considered each dual-core UltraSparc IV to be two processors for licensing purposes. That meant W.L. Gore was faced with paying double what it originally expected to pay for licensing, amounting to $100,000 per server in additional costs.

"When we first heard this, we weren't happy, of course," says Richard Sun, network systems engineer at W.L. Gore. "In all of our initial conversations with Oracle, before we purchased the licensing, we had said we were going to buy the [Sun] hardware when we were ready to implement, so we felt going to UltraSparc IV was no issue."

Analysts say W.L. Gore is not alone when it comes to being surprised by the increased software costs when updating to multiprocessor systems.

"In fact, I have spoken with clients who have been told exactly the opposite from the server vendors," says Jane Disbrow, a research director with Gartner. "They've been told, 'No, you don't have to pay double licenses. This is going to save you money in software.' I'm the one that has to give them the bad news."

While most software vendors still are feeling their way around the multi-processor issue, Oracle recently made changes in its contract language to ensure that there is no confusion. In April, the vendor changed its license agreement to redefine its processor metric to explicitly state: "For the purposes of counting the number of processors which require licensing, a multi-core chip with 'n' processors shall be counted as 'n' processors."

One reason for the confusion is that software vendors have not regarded chips with simultaneous multithreading (SMT) or Intel's Hyper-Threading technology, which make single processors work as though they were two processors, as more than a single processor.

"The reason for that is pragmatic, and that is that those two threads per core don't do anything like double the performance. It may be 10% to 25% to 30% performance improvement, but essentially SMT is a performance-enhancing technique to increase utilization of a processor core," says Gordon Haff, an analyst with Illuminata.

What isn't so obvious, some industry observers say, is what kind of performance boost users will see vs. comparable single-core systems.

Soft costs

New hardware architectures such as multi-core systems, virtualization and grid technology provide flexibility but add challenges users must cope with down the road. Consider :
Multi-core, multi-licenses: Intel and AMD plan to move to multi-core designs, and that means users across the board need to think carefully about how independent software vendors define a CPU.
Virtual machines, real costs: While per-processor licensing means users can run multiple instances of an operating system or application, many vendors still require licenses for each instance deployed in virtual environments.
Grid workloads, workload issues: Grids made up of multi-processor boxes raise issues with per-CPU licensing if workloads are moved among one-processor boxes and two-processor boxes, for example.

In addition, licensing costs associated with multi-core architectures might complicate other issues, such as buying processors for redundancy, users say.

Eric Kuzmack, IT architect at newspaper conglomerate Gannett in Silver Spring, Md., says he deploys dual-processor systems to have hardware redundancy.

"If the software vendors choose to charge us for two processors when we move to dual-core, it makes it more difficult for us to afford to put in two processors because then I'll be buying two physical packages of chips, but I'll have to pay for four processors worth of licenses, not two," he says.

Despite user concerns, most industry observers seem to agree that it's not unreasonable for software vendors to consider dual-core CPUs as two processors. Bigger issues are likely to emerge, however, as architectures move into four cores and beyond. Sun, for example, in 2006 plans to release systems based on a next-generation chip code-named Niagara, which will have processors with eight cores, each capable of running four simultaneous threads, or software instructions. Intel, which plans to release a dual-core Itanium chip next year, reportedly expects to increase the number of cores to eight and then 16 in the years ahead.

"By the time it gets beyond dual-core, the software vendors are truly going to have to think about a different way to license their products," Disbrow says. "Eventually, people won't be able to live with the software costs. The software vendors are definitely going to have to look at how they're charging for this."

How the issue can be resolved isn't clear. But analysts and users expect software vendors to change licensing plans to make them more flexible.

Sun last year began using subscription pricing for its Java Enterprise stack of middleware software, letting users pay for applications based on the number of employees, regardless of what systems the software runs on and how many services and customers it serves.

John Fowler, executive vice president of network systems at Sun, says the trend toward multi-core processors, and dynamic computing environments, in which workloads are moved among systems according to application demands, prompted the licensing change. Before last year, Sun licensed software on a per-system basis.

"What drove the change for us was the fact that customers really need to be able to move around workloads and respond to different application needs," he says. "A dynamic computing environment is not eight weeks of lawyers every time you have to renegotiate a license."

Gannett's Kuzmack complains that with virtualized systems "it takes us longer to purchase the software license than it does to provision the virtual machine."

Dan Kaberon, director of computer resource management at human resources service firm Hewitt Associates in Lincolnshire, Ill., agrees that today's software licensing is lagging behind advances in hardware and systems management. Being able to move workloads and consolidate systems on virtual machines, shouldn't result in a financial penalty on the software side, he says.

"I would just like to see a rational alignment of licensing so that if we're able to collapse multiple workloads onto a smaller number of servers that our licensing reflect that new environment and not pretend as if we were still taking advantage of all of our dedicated machines," he says. "In that case, we no longer have the upward growth and capacity opportunity and we shouldn't be penalized for it."

Tower of power

A plethora of multi-core servers and processors will hit the market in the next couple of years.  

IBM: Dual-core Power5 with multi-threading, 2004.

Sun: Dual-core UltraSparc IV with multi-threading, 2004. Niagara: Eight cores, four threads, 2006.HP: Dual-core PA-8800, 2004. Intel: All chips will be dual-core by the end of 2005. Itanium 2: Code-named Tanglewood, eight cores, 2006. AMD: Dual-core Opteron, 2005.
Must read: 11 hidden tips and tweaks for Windows 10
View Comments
Join the discussion
Be the first to comment on this article. Our Commenting Policies